人工智慧互聯的轉折點：物理層面的生態系統轉型與價值創造。

生成式人工智慧的快速發展正迫使資料中心基礎架構進行前所未有的架構改造。隨著系統瓶頸從計算處理轉向“I/O障礙”，物理層從銅纜到光纖的過渡正在加速。

本報告探討了博通和Marvell之間的戰略分歧——前者側重於漸進式優勢，後者則致力於架構轉型——並檢驗了224G SerDes的普及和超以太網聯盟（UEC）的崛起如何重塑行業標準。此外，報告還分析了台灣供應鏈如何從傳統的製造角色轉型為下一代系統的關鍵促進者和設計者。

主要亮點

• 資本轉移－網路扮演門戶角色。
  • 生成式人工智慧正在重新評估資料中心的資本支出 (CapEx) 優先順序。如果沒有網路升級，昂貴的 GPU 運算資源更有可能淪為「擱淺資產」。
• I/O限制－不是架構限制，而是物理限制。
  • 人工智慧系統正面臨關鍵的 I/O 效能障礙，因此從銅互連到光連接模組的商業化轉型至關重要。
• 建築的十字路口－演進與毀滅
  • 博通試圖透過漸進式演進來保持其橫向擴展優勢，而 Marvell 則押注於光電的橫向擴展，以突破記憶體瓶頸。這兩種願景對於人工智慧叢集的未來發展方向截然不同。
• 開放標準－重新平衡定價權
  • 224G SerDes 和 UEC主導的乙太網路正在打破單一供應商鎖定，恢復雲端服務供應商的定價權和供應鏈柔軟性。
• 台灣——從製造業到系統利用
  • 在日益加劇的地緣經濟壓力下，台灣的供應鏈正向上游工程轉移，轉向協同系統設計與先進封裝。這為隱藏平台的基礎技術奠定了基礎，從而實現了結構性利潤率的提升。

目錄

第1章：流量模式的變化重塑網路架構與標準，UEC 的出現對 InfiniBand 的主導地位構成挑戰

第2章：半導體巨頭的不同策略：博通力保其主導地位，而Marvell則大膽押注於以記憶體為中心的革命

第3章：物理限制迫使硬體重置：台灣的供應鏈從製造轉向系統實施。

第4章：TRI的觀點

The exponential growth of generative AI is forcing data center infrastructure into an unprecedented architectural overhaul. As system bottlenecks shift away from compute and toward the "I/O wall," the physical layer is accelerating its migration from copper to optics.

This report examines the strategic divergence between Broadcom and Marvell-one favoring evolutionary dominance, the other pursuing architectural disruption-and explores how the convergence of 224G SerDes and the rise of the Ultra Ethernet Consortium (UEC) are reshaping industry standards. It further analyzes how Taiwan's supply chain is emerging from traditional manufacturing roles to become a key enabler and architect of next-generation system design.

Key Highlights

• Capital Shift - Network Becomes the Gating Factor
  • Generative AI is reshaping data center CapEx priorities: without network upgrades, expensive GPU compute increasingly risks becoming stranded assets.
• I/O Wall - A Physical, Not Architectural, Constraint
  • AI system performance has hit a definitive I/O Wall, forcing a non-optional transition from copper to optical interconnects at commercial scale.
• Architectural Divide - Evolution vs. Disruption
  • Broadcom defends scale-out dominance through incremental evolution, while Marvell bets on scale-up photonics to break the memory wall-two incompatible visions of future AI clusters.
• Open Standards - Pricing Power Rebalancing
  • 224G SerDes and UEC-driven Ethernet are dismantling single-vendor lock-in, restoring pricing leverage and supply-chain flexibility to cloud service providers.
• Taiwan - From Manufacturing to System Leverage
  • Amid rising geoeconomic pressure, Taiwan's supply chain is moving upstream into system co-design and advanced packaging, enabling structural margin expansion as hidden platform enablers.

Table of Contents

1. Traffic pattern shifts are reshaping network architectures and standards as UEC emerges as a challenger to InfiniBand's Monopoly

• Table 1: AI Data Center Network Comparison: Front-End vs. Back-End
• Figure 1: Ethernet vs. InfiniBand Market Size Forecast, 2023-2028

2. Diverging strategies among chip giants: Broadcom defends its dominance, Marvell bets boldly on a memory-centric revolution

• Figure 2: Broadcom & Marvell Roadmap

3. Physical limits force a hardware reset: Taiwan's supply chain moves from manufacturing to system enablement

• Figure 3: The Evolutionary Roadmap of Interconnect Technology

4. TRI's View